Table 2.
Artificial intelligence in myopia in adults
| Tasks | Author (year) | Main predictors | AI model | Aims | Main findings |
|---|---|---|---|---|---|
| Diagnosis and detection | Lu et al., 2021[35] | Fundus images | DL | Detection of pathologic myopia | AUC - 0.979, accuracy - 0.963 |
| Tan et al., 2021[36] | Fundus images | DL | Detection of high myopia and MMD | Detection of high myopia: AUC - >0.913; detection of MMD: AUC - >0.969 | |
| Lu et al., 2021[37] | Fundus images | DL | Detection of pathologic myopia, classification of myopic maculopathy | AUC - 0.995, accuracy - 97.36%, sensitivity - 93.92%, specificity - 98.19% | |
| Choi et al., 2021[38] | OCT images | DL | Detection of high myopia | AUC - 0.86–0.99 | |
| Wan et al., 2021[39] | Fundus images | DL | Grade the risk of high myopia | AUC - 0.9968 for low-risk high myopia, AUC - 0.9964 for high-risk high myopia | |
| Li et al., 2022[40] | OCT images | DL | Detection of retinoschisis, macular hole, retinal detachment, mCNV | AUC - 0.961–0.999, sensitivity and specificity - >90% | |
| Tang et al., 2022[41] | Fundus images | DL | Grade myopic maculopathy, diagnose pathologic myopia, identify and segment myopia-related lesions | Grading accuracy - 0.9370, diagnosing pathologic myopia - 0.9980, segmentation model F1 values - 0.80–0.95 | |
| Hemelings et al., 2021[42] | Fundus images | DL | Detection of pathologic myopia; fovea localisation; segmentation of optic disc, retinal atrophy and retinal detachment | Detection of pathologic myopia: AUC - 0.9867; foveal localisation: 58.27 pixels | |
| Rauf et al., 2021[43] | Fundus images | DL | Detection of pathologic myopia | AUC - 0.9845, accuracy - 95% | |
| Du et al., 2021[44] | Fundus images | DL | Detection of pathologic myopia and myopic maculopathy (diffuse atrophy, patchy atrophy, macular atrophy, mCNV) | Diffuse atrophy AUC - 0.970, sensitivity - 84.44%; patchy atrophy AUC - 0.978, sensitivity - 87.22%; macular atrophy AUC - 0.982, sensitivity - 85.10%; mCNV AUC - 0.881, sensitivity - 37.07% | |
| Du et al., 2021[45] | OCT images | DL | Detection of myopic maculopathy | mCNV AUC - 0.985; MTM AUC - 0.946; DSM AUC - 0.978 | |
| Sogawa et al., 2020[46] | OCT images | DL | Detection of myopic macular lesions (mCNV, retinoschisis) | AUC - 0.970, sensitivity - 90.6%, specificity - 94.2% | |
| Ye et al., 2021[47] | OCT images | DL | Detection of myopic maculopathy | AUC - 0.927–0.974 | |
| Prediction | Varadarajan et al., 2018[48] | Fundus images | DL | Estimate refractive error | MAE - 0.56–0.91 diopters |
| Yoo et al., 2022[49] | Posterior segment optical coherence tomography images | DL | Estimate uncorrected refractive error; detect high myopia | SE prediction: MAE 2.66 diopters; detect high myopia: AUC - 0.813, accuracy - 71.4% | |
| Treatment | Shen et al., 2023[50] | ICL size, ACD, pupil size, ACA, CT, AL, etc. | ML | Predict the vault and the EVO-ICL size | Random forest R2=0.315, accuracy=0.828, AUC=0.765 |
| Kim et al., 2022[51] | Fundus photography, preoperative ACD, planned ablation thickness, age, preoperative CCT | ML | Identify high-risk patients for refractive regression | Combined model AUC=0.753, single model AUC=0.673 |
DL=Deep learning, ML=Machine learning, AUC=Area under the receiver operating characteristic curve, MMD=Myopic macular degeneration, OCT=Optical coherence tomography, mCNV=Myopia choroidal neovascularization, MTM=Myopic tractional maculopathy, DSM=Dome-shaped macula, MAE=Mean absolute error, ACD=Anterior chamber depth, CCT=Central corneal thickness, ACA=Anterior chamber angle, CT=Corneal thickness, AL=Axial length, AI=Artificial intelligence, ICL=Implantable collamer lens